Claude Tag in Slack, Sakana Fugu, Mistral OCR 4, Codex Remote

What shipped

Anthropic

Introducing Claude Tag

Claude joins Slack channels as a team member that anyone can tag to delegate work, in beta for Team and Enterprise, running on Opus 4.8.

Sakana AI

Sakana Fugu

General availability on June 22 of a multi-agent orchestration system sold as a single model, with Fugu Ultra as the quality tier.

Mistral AI

Mistral OCR 4

A document parser returning text, bounding boxes and confidence scores across 170 languages, at $4 per 1,000 pages.

Google DeepMind

Computer use in Gemini 3.5 Flash

June 24 makes computer use a built-in tool in Gemini 3.5 Flash, with optional confirmation steps and prompt-injection stops for enterprises.

OpenAI

Codex Remote

Generally available on June 25, letting the mobile app run Codex on a connected Mac or Windows host.

Anthropic

Claude Platform release notes, June 26

API rate limits rose and usage tiers were consolidated into Start, Build and Scale.

arXiv

Qwen-AgentWorld: Language World Models for General Agents

Language-based world models for training general agents, first on the weekly papers list.

arXiv

Wan-Streamer v0.1

End-to-end real-time interactive foundation models for video.

No new frontier model this week. It was a week of agents moving into the places where teams already talk.

Claude Tag

Anthropic introduced Claude Tag on June 23. Claude joins the channels an admin picks, can be connected to tools, data and codebases, and anyone in the channel can tag it to hand off a task. It replaces the earlier Claude in Slack app, runs on Opus 4.8, and launches as a beta for Team and Enterprise customers. Admins get token spend limits at both the organization and channel level, and there is an introductory launch credit.

I like the shape of this. One shared Claude per channel means a teammate can pick up a thread where someone else stopped, with the history right there.

Sakana Fugu and the orchestration idea

Sakana AI made Fugu generally available on June 22. It sells a multi-agent orchestration system as if it were one model behind an OpenAI-compatible API, with Fugu Ultra aimed at maximum quality on hard multi-step problems. The beta opened in April. Sakana's claims about how it compares to the top labs are its own, and I have not seen outside tests.

Documents and desktops

Mistral released OCR 4 on June 23. It returns extracted text with bounding boxes, block types such as tables, equations and signatures, and inline confidence scores, across 170 languages. Mistral reports an 85.20 on OlmOCRBench and a 72% human-preference win rate on 600-plus documents, with the benchmark caveats it states itself. The API costs $4 per 1,000 pages, $2 with batch.

Google made computer use a built-in tool in Gemini 3.5 Flash on June 24. Enterprises can require user confirmation for sensitive actions and can stop a task when an indirect prompt injection is detected. OpenAI's Codex Remote reached general availability on June 25. On June 26 Anthropic raised API rate limits and folded usage tiers into three: Start, Build and Scale.

Papers

Qwen-AgentWorld, on language world models for general agents, ranked first on Hugging Face. Wan-Streamer v0.1 aimed at real-time interactive video.

What I'd watch next: Anthropic's own Sonnet and Haiku lines, since the new rate-limit note says their limits now match Opus at every tier.